Employing integrated nested laplace approximation with stochastic partial differential equation (INLA-SPDE) modeling for sub-seasonal flood prediction using a globally aggregated disaster database

Global flood reporting has generally improved over the last century, particularly as global disaster
databases have started archiving and aggregating flood events. While these disaster databases,
including EM-DAT, DFO, HANZE, and UNDRR, are helpful tools for flood analysis, they are
often incomplete in their reporting. Therefore, aggregating them to form a more complete database
is critical. It is equally important when training a disaster-level flood event model using these
databases to account for systemic inequality in flood reporting. Here, we present a three-phase
INLA-SPDE flood prediction model which determines relevant climate signals from documented
flood events, determines which countries report floods the most consistently relative to the climate
signals, and projects global latent flood risk on a monthly 0.25-degree scale using the best reporting
countries for calibration. The result is a spatial risk field which ranks relative risk of disaster-level
flood events based on climate conditions one month in advance (sub-seasonal timescale).
Predictions at the 0.25-degree level perform well at relative risk ranking (AUC = 81.7 %) but
returns many false positives due to the complexity and rarity of these events (Precision = 6.12 %).
However, when aggregating predictions to the country level, this issue is minimized (Precision =
49.0 %), meaning country level predictions may be used to determine monthly likelihood of a flood
event occurring, while cell level predictions may be used to determine high risk locations within
the country. Combined, this modeling methodology allows for prediction of floods beyond the
capabilities of normal disaster database models by separating flood reporting biases from latent
climate signals as much as possible.